AI-Powered CSR Funding Access for Grassroots NGOs
India mandates qualifying companies to spend 2% of average net profits on CSR (Section 135, Companies Act 2013). In FY 2024-25 alone, 301 major companies spent ₹17,742 Crores on CSR activities (CSRBOX ICOR 2025). The total annual CSR pool is estimated at ~₹38,000 Crores.
Yet 3.7 million+ registered NGOs (DARPAN Portal) compete for this funding - and the vast majority lose out because they can't afford compliance teams, grant writers, or legal advisors.
NidhiAI automates the entire CSR funding lifecycle - from document verification to grant discovery to proposal generation - using a multi-agent AI system built entirely on AWS.
🔗 Live Demo: nidhi-ai.vercel.app
| Step | What Happens | AWS Service |
|---|---|---|
| 1. Upload Documents | NGO uploads 12A, 80G, CSR-1 certificates | S3 |
| 2. Compliance Check | AI extracts fields via OCR, validates dates and registration numbers | Amazon Textract + Bedrock Agent |
| 3. Grant Discovery | Semantic search matches NGO profile to corporate CSR programs | Bedrock Knowledge Bases + OpenSearch Serverless |
| 4. Proposal Drafting | AI writes a formal 5-page grant proposal, livestreamed to screen | Bedrock Agent + RAG |
| 5. Impact Reports | Generates quarterly donor reports with fund utilization and outcomes | Bedrock Agent (Amazon Nova Lite) |
| 6. CSR Chatbot | Answers questions about Section 135, Schedule VII, FCRA, 12A/80G | Bedrock Knowledge Base |
What takes 30 days and ₹50,000+ manually now takes 10 minutes at near-zero cost.
┌─────────────────────────────────────────────┐
│ Amazon Bedrock Supervisor │
│ (Multi-Agent Orchestration) │
└──────┬──────┬──────────┬──────────┬────────┘
│ │ │ │
┌──────────┘ │ │ └──────────┐
▼ ▼ ▼ ▼
┌──────────────┐ ┌────────────┐ ┌──────────────┐ ┌───────────┐
│ Compliance │ │ Grant Scout│ │ Proposal │ │ Impact │
│ Agent │ │ Agent │ │ Agent │ │ Agent │
└──────┬───────┘ └─────┬──────┘ └──────┬───────┘ └─────┬─────┘
│ │ │ │
┌──────▼───────┐ ┌─────▼──────┐ ┌──────▼───────┐ ┌────▼──────┐
│ Textract │ │ KB: CSR │ │ KB: Proposal │ │ Bedrock │
│ (OCR) │ │ Opps (OS) │ │ Templates │ │ (Nova) │
└──────────────┘ └────────────┘ └──────────────┘ └───────────┘
Frontend: Next.js 16 (Vercel) │ Auth: Amazon Cognito │ Storage: S3 + DynamoDB
Region: ap-south-1 (Mumbai) │ Compute: Lambda (x4) │ Vector DB: OpenSearch
All orchestrated by the Bedrock Supervisor Pattern - the Supervisor agent routes user requests to the correct sub-agent, chains multi-step workflows, and synthesizes responses.
| Service | Purpose |
|---|---|
| Amazon Bedrock Agents | Multi-agent orchestration using Supervisor Pattern (1 supervisor + 4 sub-agents) |
| Amazon Bedrock Knowledge Bases | 3 RAG knowledge bases - CSR laws, corporate CSR opportunities, proposal templates |
| Anthropic Claude 3.5 Sonnet (via Bedrock) | Primary foundation model for compliance reasoning and proposal generation |
| Amazon Nova Lite (via Bedrock) | Cost-effective model for grant matching and impact reports |
| Amazon Titan Text Embeddings V2 | Text-to-vector embeddings for semantic search across all 3 knowledge bases |
| Amazon Textract | OCR + form extraction on scanned government certificates (12A, 80G, CSR-1) |
| Amazon OpenSearch Serverless | Vector database backing all Bedrock Knowledge Bases |
| AWS Lambda | 4 Python 3.12 functions - one per agent action group |
| Amazon API Gateway | REST API for frontend-to-backend communication |
| Amazon S3 | Document storage (uploads, generated PDFs, KB source data) |
| Amazon DynamoDB | NGO profiles, compliance status, proposal metadata |
| Amazon Cognito | User authentication and session management |
| Amazon CloudWatch | Agent trace logging and Lambda monitoring |
| AWS IAM | Least-privilege execution roles for all services |
Region: ap-south-1 (Mumbai) - data residency compliance for Indian NGOs.
-
Real-time Streaming: Proposal and report generation use server-side streaming - the AI output appears on screen as the model writes, so users see results immediately instead of waiting 15-20 seconds for a complete response.
-
Agent Trace Panel: A built-in UI panel shows the Supervisor's decision-making live - which sub-agent was invoked, what Knowledge Base was queried, and what data was returned. Full transparency into the multi-agent orchestration.
-
100% Serverless: Zero infrastructure to manage. Scales from 0 to millions of NGOs automatically. Costs nothing when idle.
-
3 Specialized Knowledge Bases: Each KB targets a different domain - CSR legislation, corporate CSR filings, and winning proposal templates - backed by Titan Embeddings V2 and OpenSearch Serverless.
NidhiAi/
├── backend/
│ ├── api/ # API Gateway Lambda handler
│ ├── lambdas/
│ │ ├── scan_documents/ # Compliance Agent → Textract OCR
│ │ ├── match_grants/ # Grant Scout Agent → KB search
│ │ ├── generate_pdf/ # Proposal Agent → PDF generation
│ │ └── generate_report/ # Impact Agent → report generation
│ └── openapi/ # OpenAPI specs for agent action groups
├── frontend/
│ └── src/
│ ├── app/ # Next.js 16 pages (dashboard, upload, grants, proposals, reports, chatbot)
│ ├── components/ # Sidebar, AgentTrace, GrantCard, ThemeToggle
│ └── lib/ # API client, auth helpers
├── data/
│ ├── kb_csr_laws/ # Knowledge Base: Indian CSR legislation
│ ├── kb_csr_opportunities/ # Knowledge Base: Corporate CSR programs
│ └── kb_winning_proposals/ # Knowledge Base: Sample grant proposals
├── infra/ # AWS infrastructure setup scripts
├── agent_system_prompts.md # Production system prompts for all 5 Bedrock Agents
├── design.md # Technical design document
└── requirements.md # Product requirements specification
- Node.js 18+
- AWS CLI configured with
ap-south-1region - AWS services deployed (Bedrock Agents, Lambda, API Gateway, S3, DynamoDB, Cognito)
cd frontend
npm install
npm run devCreate frontend/.env.local:
NEXT_PUBLIC_API_URL=<your-api-gateway-url>
NEXT_PUBLIC_COGNITO_USER_POOL_ID=<your-pool-id>
NEXT_PUBLIC_COGNITO_CLIENT_ID=<your-client-id>
NEXT_PUBLIC_SUPERVISOR_AGENT_ID=<your-agent-id>
| Component | Monthly Cost |
|---|---|
| Lambda / API Gateway / DynamoDB | AWS Free Tier |
| Bedrock model inference | ~₹800 |
| S3 storage | < ₹50 |
| OpenSearch Serverless | Minimum capacity |
| Total prototype spend to date | < ₹8,000 |
| Projected production (100 NGOs) | ~₹17,000/month |
5,000x cost reduction per proposal compared to traditional grant writing (₹50,000+ vs < ₹10).
| Layer | Technology |
|---|---|
| Frontend | Next.js 16, React 19, TailwindCSS |
| Backend | AWS Lambda (Python 3.12) |
| AI/ML | Amazon Bedrock Agents, Textract, Titan Embeddings V2 |
| Models | Claude 4.6 Sonnet, Amazon Nova Lite |
| Database | DynamoDB, OpenSearch Serverless |
| Auth | Amazon Cognito |
| Hosting | Vercel (frontend), AWS (backend) |
Solo Developer
Built for AI for Bharat Hackathon 2026 · Track 03: AI for Communities, Access & Public Impact